How to Write AI Video Prompts Like a Filmmaker (Veo 3, Sora, Kling, Runway)
AI video models were trained on millions of clips captioned in the language of filmmaking. That’s the entire secret of good prompting: people who describe shots the way a director describes shots get dramatically better results — not because the model rewards jargon, but because “slow push-in on a close-up, backlit, anamorphic” precisely matches patterns it has seen thousands of times, while “camera slowly gets closer and it looks cool and cinematic” matches almost nothing.
This guide comes from the production side. We use these models in real film and video work, and this is the structure we actually brief them with. (Want it automated? Our AI video prompt generator builds prompts with this exact structure and formats them per model.)
The six-part shot structure
Every reliable video prompt answers six questions, in roughly this order:
- Subject & action — who/what, doing what. Concrete beats abstract: “an old fisherman coiling rope on a wet dock” not “a man working”.
- Setting & time — where, when, weather. This drives 80% of the image’s texture.
- Framing — shot size and angle: wide / medium / close-up / extreme close-up, low angle, aerial, POV.
- Camera movement — one movement, named properly: push-in, pull-back, tracking, orbit, crane, handheld, dolly zoom, FPV flythrough. One — models garble stacked moves.
- Light — the single highest-leverage word group: golden hour, blue hour, overcast soft light, neon practicals, candlelight, low-key, backlit silhouette, volumetric rays.
- Style & mood — film stock or genre reference plus emotional temperature: “vintage 16mm grain, melancholic”, “high-end commercial look, energetic”.
A complete example assembled from those parts:
Medium tracking shot following an old fisherman coiling rope on a wet dock at dawn, thick fog over the harbor. Shot on 35mm lens. Soft golden hour light breaking through the fog. Cinematic film still, photorealistic. The mood is calm and meditative.
What each model wants
Veo 3 (Google). The long-form champion — feed it full cinematic paragraphs of 60–120 words, and use its unique superpower: native audio. Dialogue in quotes, sound effects, ambience — write them into the prompt (“Audio: waves against the hull, distant gulls, he mutters: ‘Storm’s coming.’”). Veo follows detailed multi-sentence direction better than any other current model.
Sora (OpenAI). Loves dense visual description and handles complex staging — multiple subjects, physical interactions, reflections. Keep the prose rich but visual; it has no audio channel to describe.
Kling. Prefers concision. Lead with subject and a strong motion verb, add a few comma-separated aesthetic tags, and use the separate negative prompt field (blurry, distorted faces, extra limbs, text, watermark) — Kling actually honors it. Over-long prompts visibly degrade its motion.
Runway Gen-4. The official guidance is explicit: short prompts, positive phrasing only (say what you want, never “no camera shake”), and put the camera instruction first — “Slow push-in: …”. Gen-4 is the strongest image-to-video tool of the four; when you have a still frame you love, prompt only the motion and let the image carry the look.
Seedance (ByteDance). The community favorite for narrative and UGC-style work, and the model with the clearest official recipe: subject/reference → action & story beat → camera language → atmosphere & sound. Its killer feature is @image references — upload stills and write “based on @image1, the man walks down the corridor from @image2…” to lock character and location consistency across shots, which is the hardest problem in AI filmmaking. Seedance takes Chinese and English prompts equally well, and rewards describing one primary action with supporting environmental motion rather than several competing movements.
The five mistakes that ruin most AI video prompts
- No camera decision. The #1 miss. If you don’t choose framing and movement, the model chooses — badly, and differently every generation.
- Stacked movements. “The camera orbits, then pushes in, then tilts up” produces soup. One shot, one move; cut between generations like an editor would.
- Vague light. “Beautiful lighting” is noise. “Backlit by low sun, lens flare, long shadows” is a picture.
- Negative phrasing on models that ignore it. Veo/Sora/Runway: describe the presence of what you want, not the absence of what you don’t.
- Prompting a scene instead of a shot. Models generate seconds, not sequences. Break your scene into shots the way a shot list would, and generate them separately — consistency tricks (same style block, same character description verbatim) do the rest.
Camera vocabulary cheat sheet
Steal freely — these are the terms the models know best:
- Shot sizes: extreme wide, wide, medium, medium close-up, close-up, extreme close-up, POV, over-the-shoulder
- Angles: low angle, high angle, dutch angle, aerial, top-down
- Moves: push-in, pull-back, tracking/follow, orbit/arc, crane up, pedestal, whip pan, dolly zoom (Vertigo effect), handheld, FPV drone
- Lenses: 16mm wide, 35mm, 50mm natural, 85mm shallow depth of field, macro, anamorphic (oval bokeh, horizontal flares), fisheye
- Light: golden hour, blue hour, overcast, hard sun, neon practicals, candlelight, moonlight, low-key, high-key, backlit, volumetric/god rays
The words cost nothing — the decisions are the craft. Decide like a director, phrase like a caption, and generate one shot at a time. When you want the assembly automated, the prompt generator is free and runs in your browser.